godok.id

(for continuous variables) of one minus the survivor function evaluated at the variablemeans for the specified time period.4To gain a better understanding of the effects of various types of modifications onloan performance – an issue of heightened policy interest in the current economicenvironment – we estimate four regression specifications that differ by the modificationfeatures that they include. While all specifications include a HAMP indicator, the firstone (M1) does not include any other modification features; the second one (M2) adds thechange in monthly mortgage payment; the third one (M3) replaces the change in monthlymortgage payment with changes in individual loan terms including the change in loanbalance, the change in interest rate, and a term extension indicator; and the last one (M4)includes both the change in monthly mortgage payment and the changes in individualloan terms. Thus, the first regression captures a more inclusive effect of HAMP on loanperformance, but does not distinguish between effects that may be due to differences inthe magnitude of payment reductions and individual term changes between HAMP andnon-HAMP modifications, and effects that may be due to differences in program design.Differences in program design may include, for example, HAMP-specific features suchas pay-for-performance to borrowers, a requirement that borrowers work with HUDapproved counselors to reduce their debt below 55 percent (if post-modification back-enddebt-to-income (DTI) is greater than or equal to 55%), and the specific order of thewaterfall.5 While HAMP-specific eligibility criteria such as requirements that the4The cumulative probability of delinquency over a T-month period-at-risk is 1-Si(T), where Si(T) is thesurvivor function over the T-month period. In the discrete time framework of our model, Si(T) (1-Pi1) (1Pi2) (1-PiT).5Another program design feature of HAMP, the requirement of a trial period prior to the borrower beinggranted a permanent modification, has been adopted by many servicers for their proprietary modification

borrower be an owner-occupant and that the current unpaid loan balances be withinconforming loan limits also could be considered program design differences, ourregressions include specific controls for such features.6 Other distinct features of HAMP,such as the DTI eligibility criterion that qualifies only borrowers who had a front-endDTI of more than 31% at loan origination, and the requirements that this front-end DTIbe reduced to 31% and that the resulting loan must pass an NPV test, tend to result in alarger reduction in monthly payment for those borrowers who receive a HAMPmodification. (by comparison, proprietary modifications may be granted to borrowerswith original front-end ratios below 31%, but whose payment problems are due toexcessive back-end debt, and may also often result in a front-end ratio greater than 31%in order to pass NPV).The second, third, and fourth regressions help distinguish between the programdesign effects and those due instead to the magnitude of payment reductions andindividual term changes. The last regression also tests whether changes in individual loanterms have an impact on loan performance beyond any effects that would occur throughpayment changes.programs since the enactment of HAMP in 2009, and thus it is less likely to be responsible for anydifferences in redefault rates between HAMP and non-HAMP modifications in our data.6Specifically, we include a dummy variable that is equal to 1 for owner-occupied properties and 0otherwise, and the current unpaid loan balance in log terms. In preliminary work we also includedadditional indicators of HAMP eligibility such as property structure (1-4 family vs. multi-family) and adummy variable equal to 1 if loan balance at modification time was below the HAMP limit; however, thesevariables had very low statistical significance, likely due to the lack of variation of our sample across thesedimensions (e.g., 99% of the observations corresponded to 1-4 family properties and 98% of theobservations had a loan balance below the HAMP limit), and thus were excluded from the final regressions.In addition, we experimented with a single indicator that captured the joint HAMP eligibility under the loanlimit, owner occupancy, and property structure criteria. This indicator also had very low significance leveland its inclusion left the results virtually unchanged. Results from these alternative specifications areavailable upon request from the authors.

In additional specifications, we explore variation in the effects over time, and testwhether the effects of modification features such as payment change, balance change,rate change, and term extension vary with the borrower’s credit score (FICO) and loan tovalue (LTV) levels. Temporal variations in any performance differential between HAMPand non-HAMP modifications may occur as a result of changes in the structure ofproprietary loan modifications (perhaps in part due to the advent of HAMP itself) as wellas to changes in HAMP rules (e.g., such as those in Supplemental Directive 10-01 fromJune 2010 including new rules regarding documentation requirements and amendments topolicies and procedures related to borrower outreach and communication).To explore these temporal dynamics, we supplement model M1 with twovariables that capture the pre- and post-HAMP enactment time trends, a post-HAMPenactment dummy variable, and an interaction between the HAMP indicator and the postHAMP enactment time trend.7 The time trend and post-HAMP dummy variables describethe comparative loan performance of older and newer vintages of proprietarymodifications, allowing for a direct comparison of the performance of the pre-HAMP andpost-HAMP proprietary modifications. The HAMP indicator and its interaction with thepost-HAMP trend capture temporal variations in the differential performance of HAMPmodifications versus proprietary modifications granted in the post-HAMP period.To test whether the effects of modification features vary with the FICO and LTVlevels, we extend models M2 through M4 to include interactions between the relevant7The post-HAMP enactment period is assumed to start in September 2009 when the first permanentHAMP modifications were completed, according to our Mortgage Metrics data extract for New York City.Thus, the post-HAMP time trend is equal to 0 if the modification was completed prior to September 2009,is equal to 1 if the modification was completed in September 2009, is equal to 2 if the modification wascompleted in October 2009, etc. The pre-HAMP time trend is equal to 0 if the modification was completedin August 2009 or later, is equal to -1 if the modification was completed in July 2009, is equal to -2 if themodification was completed in June 2009, etc.

modification changes and indicators for the lowest FICO category (FICO less than 560)and for the largest LTV category (LTV greater than 120 percent), respectively.4. Data DescriptionTo investigate the determinants of the performance of modified loans, we analyzeperformance between January 2008 and November 2010 for all first lien mortgagesoriginated in New York City from 2004 to 2008 and still active as of January 1, 2008 inthe OCC Mortgage Metrics database. OCC Mortgage Metrics provides loan-level data onloan characteristics and performance, including detailed information about loanmodifications, for residential mortgages serviced by selected national banks and federalsavings associations. The database includes loans serviced by 9 large mortgage servicerscovering 63 percent of all mortgages outstanding in the United States, and includes alltypes of mortgages serviced, including both prime and subprime mortgages (OCC andOTS, 2011).8 Nationally, the loans in the OCC Mortgage Metrics dataset represent a largeshare of the overall mortgage industry, but they do not represent a statistically randomsample of all mortgage loans. For example, only the largest servicers are included in theOCC Mortgage Metrics, and a large majority of the included servicers are national banks.Thus, the characteristics of these loans may differ from the overall population ofmortgages in the United States. For example, subprime mortgages are underrepresentedand conforming loans sold to the GSEs are overrepresented in the OCC MortgageMetrics data (U.S. Department of Treasury, 2008).8The number of servicers in the OCC Mortgage Metrics has varied over time since the onset of the datacollection in 2007, primarily due to mergers and acquisitions among the initial servicers that provided thedata. As of 2011, the servicers in the OCC Mortgage Metrics include 8 national banks and one thrift withthe largest mortgage-servicing portfolios among national banks and thrifts (OCC and OTS, 2011).

An observation in the data set is a loan in a given month. Although we look at allloans originated between 2004 and 2008, monthly performance history for those loans isonly available from January 2008 through November 2010. If a loan was originated in2004 and went through foreclosure proceedings in 2007, therefore, we will never see thatloan. Although OCC Mortgage Metrics provides detailed information on borrowercharacteristics, loan terms, payment history and modifications, it contains no informationon borrower race or gender and provides little information about property orneighborhood characteristics. We therefore supplement the loan level data withinformation from multiple sources.To match loan level information from the OCC Mortgage Metrics database toother sources, we relied on mortgage deeds contained within the New York CityDepartment of Finance’s Automated City Register Information System (ACRIS). Using ahierarchical matching algorithm, we were able to match 65 percent of the loans in theOCC Mortgage Metrics database back to the deeds records, which thus gave us the exact

location of the mortgaged property.9 This 65 percent sample is not significantly differentfrom the full universe in terms of the loan and borrower characteristics that we use in theanalyses below.After we had a unique parcel identifier matched to each loan record, we were ableto match on many other sources. First, we attach some additional borrowercharacteristics, including race and ethnicity, from Home Mortgage Disclosure Act(HMDA) data.10 Second, we merge information on whether the borrower receivedforeclosure prevention counseling or other assistance from any of the non-profitorganizations coordinated by the Center for New York City Neighborhoods (CNYCN).119Our procedure for matching OCC Mortgage Metrics to ACRIS is similar to the method used by Chan etal. (2010) to match LoanPerformance to ACRIS. Our data from ACRIS do not include Staten Island andthus we had to drop this borough from our analysis. We merged OCC Mortgage Metrics loans to ACRISmortgage deeds using three common fields: origination or deed date, loan amount and zip code, using sixstages of hierarchical matching. At the end of each stage, loans and deeds that uniquely matched each otherwere set aside and considered matched, while all other loans and deeds enter the next stage. Stage 1matched loans and deeds on the raw values of date, loan amount and zip code. Stage 2 matched theremaining loans and deeds on the raw values of date and zip code, and the loan amount rounded to 1,000.Stage 3 matched on the raw values of date and zip code, and the loan amount rounded to 10,000. Stage 4matched on the raw values of zip code and loan amount, and allowed dates to differ by up to 60 days. Stage5 matched on the raw value of zip code, loan amount rounded to 1,000, and allowed dates to differ by upto 60 days. Stage 6 matched on the raw value of zip code, loan amount rounded to 10,000, and alloweddates to differ by up to 60 days. We believe it is valid to introduce a 60-day window because in ACRIS,there may be administrative lags in the recording of the deeds data. The chance of false positive matching islow because we are matching loans to the full universe of deed records, and only considering uniquematches. The relatively low match rate of 65 percent is due to the fact that we were unable to match loansmade on coop units in the OCC Mortgage Metrics data to ACRIS deeds because coop mortgages arerecorded differently in ACRIS and do not list a loan amount. During our study period, 28 percent ofresidential property sales in the four boroughs studied were coops. Further, our match rate was lowest (44percent) in Manhattan where 48 percent of sales during the study period were of coop units. This evidencesuggests that had we been able to exclude coop loans from our original OCC Mortgage Metrics datasetprior to matching to ACRIS, our final match rate would have been much higher (around 90 percent).10We merged HMDA records to ACRIS deeds based on date, loan amount and census tract, using the samesix stage hierarchical matching technique as for the OCC Mortgage Metrics-ACRIS match. We then pairedeach of the OCC Mortgage Metrics records with HMDA records based on the unique deed identificationnumber from ACRIS. In the end, we were able to match 73 percent of the OCC Mortgage Metrics-ACRISmatched loans (or 48 percent of all OCC Mortgage Metrics loans) to the HMDA records. While otherresearchers have matched loan level data (such as OCC Mortgage Metrics) directly to HMDA by using thezip code as a common geographic identifier, our matching strategy is likely more reliable as it uses a moreprecise common geographical identifier (census tract).11CNYCN is a non-profit organization, funded by grants from government, foundations, and financialinstitutions, to coordinate foreclosure counseling, education, and legal services from a variety of non-profit

Third, we merge in repeat sales house price indices the Furman Center for Real Estateand Urban Policy compiles to track appreciation in 56 different community districts ofNew York City.12 Fourth, we link information on the demographic characteristics ofcensus tracts using the 2000 Census. Finally, we add the rate of mortgage foreclosurenotices (lis pendens) at the census tract level.13When available, we matched data at the observation level to show informationabout the specific property being studied. When observation level data was not available(e.g., educational attainment) or was not appropriate (e.g., 6 month prior neighborhood lispendens rate) we used neighborhood level data instead. We define neighborhood as acensus tract, the smallest geographic level available, whenever possible. However, forseveral variables – specifically, the unemployment rate and the rate of house priceappreciation – census tract data was not available, so we had to use community districtlevel data.14 To illustrate the relative size of each jurisdiction, Figure 1 shows census tractboundaries, community district boundaries and lis pendens filed in the four boroughs ofNew York City in 2009.15providers throughout New York City to homeowners and tenants at risk of losing their home to foreclosure.CNYCN directs borrowers facing trouble with their mortgages who call 311 or CNYCN directly to localforeclosure counseling or legal services. Each of its partner organizations then reports back to CNYCN onwhich borrowers received foreclosure prevention counseling or legal services. One of the co-authors, VickiBeen, serves on the Board of Directors for CNYCN.12See Armstrong et al. (2009) for a description. We transform quarterly indices into monthly series bylinear interpolation.13The lis pendens are from Furman Center’s calculations based on data from Public Data Corporation. Therate is computed as the number of lis pendens per 1000 housing units recorded over the 6-month periodpreceding the month of loan performance.14Community districts are political units unique to New York City. Each of the 59 community districts hasa Community Board that makes non-binding recommendations about applications for zoning changes andother land use proposals, and recommends budget priorities.15For readability purposes, we do not show zip code boundaries in this map. We note however that thetypical zip code size, both in terms of area and population, is larger than the typical census tract size butsmaller than the typical community district size.

4.1 Descriptive StatisticsTable 1 presents descriptive statistics for the dataset used in the estimation,organized in six panels: A – delinquency rates; B – modification features; C– loan characteristics; D – borrower and property characteristics; E – neighborhoodcharacteristics; and F – servicer characteristics. Panel A shows that nearly 30 percent ofthe modified loans in our data became seriously delinquent following modification. Amore informative description of the performance of modified loans is provided by theKaplan-Meyer survival graph in Figure 2A. The survival graph plots, over time sincemodification, the fraction of the modified loans that have “survived”, in that they havenot yet redefaulted. Given our definition of redefault as the payment becoming 60 dayspast due, the first month that a loan is “at risk” is in the second month after modification,and the origin of the survival plot in Figure 2A corresponds to the first month followingmodification. Notice that, starting in the second month after modification, there is asteady transition of loans into serious delinquency with the pace diminishing beyond the15th month following modification. The survival rate one year after modification is justbelow 60 percent. Figure 2B shows sharp differences in survival rates between the loansthat received HAMP modifications and those that received proprietary modifications. Forexample, the survival rate of HAMP loans one year after modification is over 30percentage points higher than the survival rate of non-HAMP loans.Panel B of Table 1 presents descriptive statistics for the types of the modificationsin our sample. One third of the loans received HAMP modifications. The modificationprocess resulted in payment reductions for most – but not all - loans. While over 80percent of the modifications resulted in payment reductions, almost 7 percent resulted in

payment increases and nearly 4 percent produced no payment change. On average, themortgage payment was reduced by 28 percent. A majority of the modifications resulted inhigher balances, while only about 10 percent resulted in lower balances and almost 15percent produced no balance change. On average, the balance was increased by 2.6percent. The prevalence of balance increases is not surprising given that capitalization –the addition of arrearages to the loan balance – is a frequent component of themodifications in our data, whereas principal write-down is very rarely used.16 Over 75percent of the modifications resulted in a decrease in interest rates, and the ratereductions were substantial -- 2.8 percentage points, on average. Approximately 45percent of the modifications included term extensions, however the actual size of the termchange was largely missing in our data and thus we could not use this information in ouranalysis. Overall, these patterns suggest that servicers aim to make the loans moreaffordable while minimizing losses in the underlying principal.Panel C presents descriptive statistics for the characteristics of the loans in ourdataset. Our dataset covers a range of loan products. Of the 6,541 modified loans in ourdataset: there is a nearly even split between prime and non-prime loans;17 57 percent havefixed interest rates while the remainder have adjustable rate mortgages; 14 percent wereinterest only at origination and 79 percent are conventional mortgages. Our sample alsoincludes a mix of loans that were privately securitized, bought by the GSEs and held inportfolio. This robust mix of loan products, uses and investors allows us to give a more16Almost 90 percent of the modifications involved capitalization whereas only about 2 percent includedprincipal write-down.17Loans are categorized as prime or non-prime based on the credit grades defined by the servicers.

complete analysis than the existing literature because our conclusions are not limited toonly one loan type or group of loans.The relative interest rate after modification for FRMs is calculated as the interestrate mi

the loan, and (3) extending the term of the loan, sometimes by adding missed payments to the principal. Generally (but not always) a combination of these modification strategies will result in a lower monthly payment for the borrower. However, some modifications have employed these tools in such a way that the monthly payment actually increased.